1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Jump Threading pass.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Scalar.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/CFG.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/LazyValueInfo.h"
25 #include "llvm/Analysis/Loads.h"
26 #include "llvm/Analysis/TargetLibraryInfo.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/IR/Metadata.h"
31 #include "llvm/IR/ValueHandle.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Local.h"
38 #include "llvm/Transforms/Utils/SSAUpdater.h"
41 #define DEBUG_TYPE "jump-threading"
43 STATISTIC(NumThreads, "Number of jumps threaded");
44 STATISTIC(NumFolds, "Number of terminators folded");
45 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
47 static cl::opt<unsigned>
48 BBDuplicateThreshold("jump-threading-threshold",
49 cl::desc("Max block size to duplicate for jump threading"),
50 cl::init(6), cl::Hidden);
53 // These are at global scope so static functions can use them too.
54 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
55 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
57 // This is used to keep track of what kind of constant we're currently hoping
59 enum ConstantPreference {
64 /// This pass performs 'jump threading', which looks at blocks that have
65 /// multiple predecessors and multiple successors. If one or more of the
66 /// predecessors of the block can be proven to always jump to one of the
67 /// successors, we forward the edge from the predecessor to the successor by
68 /// duplicating the contents of this block.
70 /// An example of when this can occur is code like this:
77 /// In this case, the unconditional branch at the end of the first if can be
78 /// revectored to the false side of the second if.
80 class JumpThreading : public FunctionPass {
81 TargetLibraryInfo *TLI;
84 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
86 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
88 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
90 unsigned BBDupThreshold;
92 // RAII helper for updating the recursion stack.
93 struct RecursionSetRemover {
94 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
95 std::pair<Value*, BasicBlock*> ThePair;
97 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
98 std::pair<Value*, BasicBlock*> P)
99 : TheSet(S), ThePair(P) { }
101 ~RecursionSetRemover() {
102 TheSet.erase(ThePair);
106 static char ID; // Pass identification
107 JumpThreading(int T = -1) : FunctionPass(ID) {
108 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
109 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
112 bool runOnFunction(Function &F) override;
114 void getAnalysisUsage(AnalysisUsage &AU) const override {
115 AU.addRequired<LazyValueInfo>();
116 AU.addPreserved<LazyValueInfo>();
117 AU.addRequired<TargetLibraryInfoWrapperPass>();
120 void FindLoopHeaders(Function &F);
121 bool ProcessBlock(BasicBlock *BB);
122 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
124 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
125 const SmallVectorImpl<BasicBlock *> &PredBBs);
127 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
128 PredValueInfo &Result,
129 ConstantPreference Preference,
130 Instruction *CxtI = nullptr);
131 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
132 ConstantPreference Preference,
133 Instruction *CxtI = nullptr);
135 bool ProcessBranchOnPHI(PHINode *PN);
136 bool ProcessBranchOnXOR(BinaryOperator *BO);
138 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
139 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
143 char JumpThreading::ID = 0;
144 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
145 "Jump Threading", false, false)
146 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
147 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
148 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
149 "Jump Threading", false, false)
151 // Public interface to the Jump Threading pass
152 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); }
154 /// runOnFunction - Top level algorithm.
156 bool JumpThreading::runOnFunction(Function &F) {
157 if (skipOptnoneFunction(F))
160 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
161 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
162 LVI = &getAnalysis<LazyValueInfo>();
164 // Remove unreachable blocks from function as they may result in infinite
165 // loop. We do threading if we found something profitable. Jump threading a
166 // branch can create other opportunities. If these opportunities form a cycle
167 // i.e. if any jump treading is undoing previous threading in the path, then
168 // we will loop forever. We take care of this issue by not jump threading for
169 // back edges. This works for normal cases but not for unreachable blocks as
170 // they may have cycle with no back edge.
171 removeUnreachableBlocks(F);
175 bool Changed, EverChanged = false;
178 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
180 // Thread all of the branches we can over this block.
181 while (ProcessBlock(BB))
186 // If the block is trivially dead, zap it. This eliminates the successor
187 // edges which simplifies the CFG.
188 if (pred_empty(BB) &&
189 BB != &BB->getParent()->getEntryBlock()) {
190 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
191 << "' with terminator: " << *BB->getTerminator() << '\n');
192 LoopHeaders.erase(BB);
199 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
201 // Can't thread an unconditional jump, but if the block is "almost
202 // empty", we can replace uses of it with uses of the successor and make
204 if (BI && BI->isUnconditional() &&
205 BB != &BB->getParent()->getEntryBlock() &&
206 // If the terminator is the only non-phi instruction, try to nuke it.
207 BB->getFirstNonPHIOrDbg()->isTerminator()) {
208 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
209 // block, we have to make sure it isn't in the LoopHeaders set. We
210 // reinsert afterward if needed.
211 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
212 BasicBlock *Succ = BI->getSuccessor(0);
214 // FIXME: It is always conservatively correct to drop the info
215 // for a block even if it doesn't get erased. This isn't totally
216 // awesome, but it allows us to use AssertingVH to prevent nasty
217 // dangling pointer issues within LazyValueInfo.
219 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
221 // If we deleted BB and BB was the header of a loop, then the
222 // successor is now the header of the loop.
226 if (ErasedFromLoopHeaders)
227 LoopHeaders.insert(BB);
230 EverChanged |= Changed;
237 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
238 /// thread across it. Stop scanning the block when passing the threshold.
239 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
240 unsigned Threshold) {
241 /// Ignore PHI nodes, these will be flattened when duplication happens.
242 BasicBlock::const_iterator I = BB->getFirstNonPHI();
244 // FIXME: THREADING will delete values that are just used to compute the
245 // branch, so they shouldn't count against the duplication cost.
247 // Sum up the cost of each instruction until we get to the terminator. Don't
248 // include the terminator because the copy won't include it.
250 for (; !isa<TerminatorInst>(I); ++I) {
252 // Stop scanning the block if we've reached the threshold.
253 if (Size > Threshold)
256 // Debugger intrinsics don't incur code size.
257 if (isa<DbgInfoIntrinsic>(I)) continue;
259 // If this is a pointer->pointer bitcast, it is free.
260 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
263 // Bail out if this instruction gives back a token type, it is not possible
264 // to duplicate it if it used outside this BB.
265 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
268 // All other instructions count for at least one unit.
271 // Calls are more expensive. If they are non-intrinsic calls, we model them
272 // as having cost of 4. If they are a non-vector intrinsic, we model them
273 // as having cost of 2 total, and if they are a vector intrinsic, we model
274 // them as having cost 1.
275 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
276 if (CI->cannotDuplicate() || CI->isConvergent())
277 // Blocks with NoDuplicate are modelled as having infinite cost, so they
278 // are never duplicated.
280 else if (!isa<IntrinsicInst>(CI))
282 else if (!CI->getType()->isVectorTy())
287 // Threading through a switch statement is particularly profitable. If this
288 // block ends in a switch, decrease its cost to make it more likely to happen.
289 if (isa<SwitchInst>(I))
290 Size = Size > 6 ? Size-6 : 0;
292 // The same holds for indirect branches, but slightly more so.
293 if (isa<IndirectBrInst>(I))
294 Size = Size > 8 ? Size-8 : 0;
299 /// FindLoopHeaders - We do not want jump threading to turn proper loop
300 /// structures into irreducible loops. Doing this breaks up the loop nesting
301 /// hierarchy and pessimizes later transformations. To prevent this from
302 /// happening, we first have to find the loop headers. Here we approximate this
303 /// by finding targets of backedges in the CFG.
305 /// Note that there definitely are cases when we want to allow threading of
306 /// edges across a loop header. For example, threading a jump from outside the
307 /// loop (the preheader) to an exit block of the loop is definitely profitable.
308 /// It is also almost always profitable to thread backedges from within the loop
309 /// to exit blocks, and is often profitable to thread backedges to other blocks
310 /// within the loop (forming a nested loop). This simple analysis is not rich
311 /// enough to track all of these properties and keep it up-to-date as the CFG
312 /// mutates, so we don't allow any of these transformations.
314 void JumpThreading::FindLoopHeaders(Function &F) {
315 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
316 FindFunctionBackedges(F, Edges);
318 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
319 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
322 /// getKnownConstant - Helper method to determine if we can thread over a
323 /// terminator with the given value as its condition, and if so what value to
324 /// use for that. What kind of value this is depends on whether we want an
325 /// integer or a block address, but an undef is always accepted.
326 /// Returns null if Val is null or not an appropriate constant.
327 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
331 // Undef is "known" enough.
332 if (UndefValue *U = dyn_cast<UndefValue>(Val))
335 if (Preference == WantBlockAddress)
336 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
338 return dyn_cast<ConstantInt>(Val);
341 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
342 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
343 /// in any of our predecessors. If so, return the known list of value and pred
344 /// BB in the result vector.
346 /// This returns true if there were any known values.
349 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
350 ConstantPreference Preference,
352 // This method walks up use-def chains recursively. Because of this, we could
353 // get into an infinite loop going around loops in the use-def chain. To
354 // prevent this, keep track of what (value, block) pairs we've already visited
355 // and terminate the search if we loop back to them
356 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
359 // An RAII help to remove this pair from the recursion set once the recursion
360 // stack pops back out again.
361 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
363 // If V is a constant, then it is known in all predecessors.
364 if (Constant *KC = getKnownConstant(V, Preference)) {
365 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
366 Result.push_back(std::make_pair(KC, *PI));
371 // If V is a non-instruction value, or an instruction in a different block,
372 // then it can't be derived from a PHI.
373 Instruction *I = dyn_cast<Instruction>(V);
374 if (!I || I->getParent() != BB) {
376 // Okay, if this is a live-in value, see if it has a known value at the end
377 // of any of our predecessors.
379 // FIXME: This should be an edge property, not a block end property.
380 /// TODO: Per PR2563, we could infer value range information about a
381 /// predecessor based on its terminator.
383 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
384 // "I" is a non-local compare-with-a-constant instruction. This would be
385 // able to handle value inequalities better, for example if the compare is
386 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
387 // Perhaps getConstantOnEdge should be smart enough to do this?
389 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
391 // If the value is known by LazyValueInfo to be a constant in a
392 // predecessor, use that information to try to thread this block.
393 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
394 if (Constant *KC = getKnownConstant(PredCst, Preference))
395 Result.push_back(std::make_pair(KC, P));
398 return !Result.empty();
401 /// If I is a PHI node, then we know the incoming values for any constants.
402 if (PHINode *PN = dyn_cast<PHINode>(I)) {
403 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
404 Value *InVal = PN->getIncomingValue(i);
405 if (Constant *KC = getKnownConstant(InVal, Preference)) {
406 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
408 Constant *CI = LVI->getConstantOnEdge(InVal,
409 PN->getIncomingBlock(i),
411 if (Constant *KC = getKnownConstant(CI, Preference))
412 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
416 return !Result.empty();
419 PredValueInfoTy LHSVals, RHSVals;
421 // Handle some boolean conditions.
422 if (I->getType()->getPrimitiveSizeInBits() == 1) {
423 assert(Preference == WantInteger && "One-bit non-integer type?");
425 // X & false -> false
426 if (I->getOpcode() == Instruction::Or ||
427 I->getOpcode() == Instruction::And) {
428 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
430 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
433 if (LHSVals.empty() && RHSVals.empty())
436 ConstantInt *InterestingVal;
437 if (I->getOpcode() == Instruction::Or)
438 InterestingVal = ConstantInt::getTrue(I->getContext());
440 InterestingVal = ConstantInt::getFalse(I->getContext());
442 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
444 // Scan for the sentinel. If we find an undef, force it to the
445 // interesting value: x|undef -> true and x&undef -> false.
446 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
447 if (LHSVals[i].first == InterestingVal ||
448 isa<UndefValue>(LHSVals[i].first)) {
449 Result.push_back(LHSVals[i]);
450 Result.back().first = InterestingVal;
451 LHSKnownBBs.insert(LHSVals[i].second);
453 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
454 if (RHSVals[i].first == InterestingVal ||
455 isa<UndefValue>(RHSVals[i].first)) {
456 // If we already inferred a value for this block on the LHS, don't
458 if (!LHSKnownBBs.count(RHSVals[i].second)) {
459 Result.push_back(RHSVals[i]);
460 Result.back().first = InterestingVal;
464 return !Result.empty();
467 // Handle the NOT form of XOR.
468 if (I->getOpcode() == Instruction::Xor &&
469 isa<ConstantInt>(I->getOperand(1)) &&
470 cast<ConstantInt>(I->getOperand(1))->isOne()) {
471 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
476 // Invert the known values.
477 for (unsigned i = 0, e = Result.size(); i != e; ++i)
478 Result[i].first = ConstantExpr::getNot(Result[i].first);
483 // Try to simplify some other binary operator values.
484 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
485 assert(Preference != WantBlockAddress
486 && "A binary operator creating a block address?");
487 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
488 PredValueInfoTy LHSVals;
489 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
492 // Try to use constant folding to simplify the binary operator.
493 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
494 Constant *V = LHSVals[i].first;
495 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
497 if (Constant *KC = getKnownConstant(Folded, WantInteger))
498 Result.push_back(std::make_pair(KC, LHSVals[i].second));
502 return !Result.empty();
505 // Handle compare with phi operand, where the PHI is defined in this block.
506 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
507 assert(Preference == WantInteger && "Compares only produce integers");
508 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
509 if (PN && PN->getParent() == BB) {
510 const DataLayout &DL = PN->getModule()->getDataLayout();
511 // We can do this simplification if any comparisons fold to true or false.
513 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
514 BasicBlock *PredBB = PN->getIncomingBlock(i);
515 Value *LHS = PN->getIncomingValue(i);
516 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
518 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
520 if (!isa<Constant>(RHS))
523 LazyValueInfo::Tristate
524 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
525 cast<Constant>(RHS), PredBB, BB,
527 if (ResT == LazyValueInfo::Unknown)
529 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
532 if (Constant *KC = getKnownConstant(Res, WantInteger))
533 Result.push_back(std::make_pair(KC, PredBB));
536 return !Result.empty();
539 // If comparing a live-in value against a constant, see if we know the
540 // live-in value on any predecessors.
541 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
542 if (!isa<Instruction>(Cmp->getOperand(0)) ||
543 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
544 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
546 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
548 // If the value is known by LazyValueInfo to be a constant in a
549 // predecessor, use that information to try to thread this block.
550 LazyValueInfo::Tristate Res =
551 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
552 RHSCst, P, BB, CxtI ? CxtI : Cmp);
553 if (Res == LazyValueInfo::Unknown)
556 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
557 Result.push_back(std::make_pair(ResC, P));
560 return !Result.empty();
563 // Try to find a constant value for the LHS of a comparison,
564 // and evaluate it statically if we can.
565 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
566 PredValueInfoTy LHSVals;
567 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
570 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
571 Constant *V = LHSVals[i].first;
572 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
574 if (Constant *KC = getKnownConstant(Folded, WantInteger))
575 Result.push_back(std::make_pair(KC, LHSVals[i].second));
578 return !Result.empty();
583 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
584 // Handle select instructions where at least one operand is a known constant
585 // and we can figure out the condition value for any predecessor block.
586 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
587 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
588 PredValueInfoTy Conds;
589 if ((TrueVal || FalseVal) &&
590 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
591 WantInteger, CxtI)) {
592 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
593 Constant *Cond = Conds[i].first;
595 // Figure out what value to use for the condition.
597 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
599 KnownCond = CI->isOne();
601 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
602 // Either operand will do, so be sure to pick the one that's a known
604 // FIXME: Do this more cleverly if both values are known constants?
605 KnownCond = (TrueVal != nullptr);
608 // See if the select has a known constant value for this predecessor.
609 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
610 Result.push_back(std::make_pair(Val, Conds[i].second));
613 return !Result.empty();
617 // If all else fails, see if LVI can figure out a constant value for us.
618 Constant *CI = LVI->getConstant(V, BB, CxtI);
619 if (Constant *KC = getKnownConstant(CI, Preference)) {
620 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
621 Result.push_back(std::make_pair(KC, *PI));
624 return !Result.empty();
629 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
630 /// in an undefined jump, decide which block is best to revector to.
632 /// Since we can pick an arbitrary destination, we pick the successor with the
633 /// fewest predecessors. This should reduce the in-degree of the others.
635 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
636 TerminatorInst *BBTerm = BB->getTerminator();
637 unsigned MinSucc = 0;
638 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
639 // Compute the successor with the minimum number of predecessors.
640 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
641 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
642 TestBB = BBTerm->getSuccessor(i);
643 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
644 if (NumPreds < MinNumPreds) {
646 MinNumPreds = NumPreds;
653 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
654 if (!BB->hasAddressTaken()) return false;
656 // If the block has its address taken, it may be a tree of dead constants
657 // hanging off of it. These shouldn't keep the block alive.
658 BlockAddress *BA = BlockAddress::get(BB);
659 BA->removeDeadConstantUsers();
660 return !BA->use_empty();
663 /// ProcessBlock - If there are any predecessors whose control can be threaded
664 /// through to a successor, transform them now.
665 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
666 // If the block is trivially dead, just return and let the caller nuke it.
667 // This simplifies other transformations.
668 if (pred_empty(BB) &&
669 BB != &BB->getParent()->getEntryBlock())
672 // If this block has a single predecessor, and if that pred has a single
673 // successor, merge the blocks. This encourages recursive jump threading
674 // because now the condition in this block can be threaded through
675 // predecessors of our predecessor block.
676 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
677 const TerminatorInst *TI = SinglePred->getTerminator();
678 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
679 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
680 // If SinglePred was a loop header, BB becomes one.
681 if (LoopHeaders.erase(SinglePred))
682 LoopHeaders.insert(BB);
684 LVI->eraseBlock(SinglePred);
685 MergeBasicBlockIntoOnlyPred(BB);
691 // What kind of constant we're looking for.
692 ConstantPreference Preference = WantInteger;
694 // Look to see if the terminator is a conditional branch, switch or indirect
695 // branch, if not we can't thread it.
697 Instruction *Terminator = BB->getTerminator();
698 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
699 // Can't thread an unconditional jump.
700 if (BI->isUnconditional()) return false;
701 Condition = BI->getCondition();
702 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
703 Condition = SI->getCondition();
704 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
705 // Can't thread indirect branch with no successors.
706 if (IB->getNumSuccessors() == 0) return false;
707 Condition = IB->getAddress()->stripPointerCasts();
708 Preference = WantBlockAddress;
710 return false; // Must be an invoke.
713 // Run constant folding to see if we can reduce the condition to a simple
715 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
717 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
719 I->replaceAllUsesWith(SimpleVal);
720 I->eraseFromParent();
721 Condition = SimpleVal;
725 // If the terminator is branching on an undef, we can pick any of the
726 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
727 if (isa<UndefValue>(Condition)) {
728 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
730 // Fold the branch/switch.
731 TerminatorInst *BBTerm = BB->getTerminator();
732 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
733 if (i == BestSucc) continue;
734 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
737 DEBUG(dbgs() << " In block '" << BB->getName()
738 << "' folding undef terminator: " << *BBTerm << '\n');
739 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
740 BBTerm->eraseFromParent();
744 // If the terminator of this block is branching on a constant, simplify the
745 // terminator to an unconditional branch. This can occur due to threading in
747 if (getKnownConstant(Condition, Preference)) {
748 DEBUG(dbgs() << " In block '" << BB->getName()
749 << "' folding terminator: " << *BB->getTerminator() << '\n');
751 ConstantFoldTerminator(BB, true);
755 Instruction *CondInst = dyn_cast<Instruction>(Condition);
757 // All the rest of our checks depend on the condition being an instruction.
759 // FIXME: Unify this with code below.
760 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
766 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
767 // If we're branching on a conditional, LVI might be able to determine
768 // it's value at the branch instruction. We only handle comparisons
769 // against a constant at this time.
770 // TODO: This should be extended to handle switches as well.
771 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
772 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
773 if (CondBr && CondConst && CondBr->isConditional()) {
774 LazyValueInfo::Tristate Ret =
775 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
777 if (Ret != LazyValueInfo::Unknown) {
778 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
779 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
780 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
781 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
782 CondBr->eraseFromParent();
783 if (CondCmp->use_empty())
784 CondCmp->eraseFromParent();
785 else if (CondCmp->getParent() == BB) {
786 // If the fact we just learned is true for all uses of the
787 // condition, replace it with a constant value
788 auto *CI = Ret == LazyValueInfo::True ?
789 ConstantInt::getTrue(CondCmp->getType()) :
790 ConstantInt::getFalse(CondCmp->getType());
791 CondCmp->replaceAllUsesWith(CI);
792 CondCmp->eraseFromParent();
798 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
802 // Check for some cases that are worth simplifying. Right now we want to look
803 // for loads that are used by a switch or by the condition for the branch. If
804 // we see one, check to see if it's partially redundant. If so, insert a PHI
805 // which can then be used to thread the values.
807 Value *SimplifyValue = CondInst;
808 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
809 if (isa<Constant>(CondCmp->getOperand(1)))
810 SimplifyValue = CondCmp->getOperand(0);
812 // TODO: There are other places where load PRE would be profitable, such as
813 // more complex comparisons.
814 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
815 if (SimplifyPartiallyRedundantLoad(LI))
819 // Handle a variety of cases where we are branching on something derived from
820 // a PHI node in the current block. If we can prove that any predecessors
821 // compute a predictable value based on a PHI node, thread those predecessors.
823 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
826 // If this is an otherwise-unfoldable branch on a phi node in the current
827 // block, see if we can simplify.
828 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
829 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
830 return ProcessBranchOnPHI(PN);
833 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
834 if (CondInst->getOpcode() == Instruction::Xor &&
835 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
836 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
839 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
840 // "(X == 4)", thread through this block.
845 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
846 /// load instruction, eliminate it by replacing it with a PHI node. This is an
847 /// important optimization that encourages jump threading, and needs to be run
848 /// interlaced with other jump threading tasks.
849 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
850 // Don't hack volatile/atomic loads.
851 if (!LI->isSimple()) return false;
853 // If the load is defined in a block with exactly one predecessor, it can't be
854 // partially redundant.
855 BasicBlock *LoadBB = LI->getParent();
856 if (LoadBB->getSinglePredecessor())
859 // If the load is defined in an EH pad, it can't be partially redundant,
860 // because the edges between the invoke and the EH pad cannot have other
861 // instructions between them.
862 if (LoadBB->isEHPad())
865 Value *LoadedPtr = LI->getOperand(0);
867 // If the loaded operand is defined in the LoadBB, it can't be available.
868 // TODO: Could do simple PHI translation, that would be fun :)
869 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
870 if (PtrOp->getParent() == LoadBB)
873 // Scan a few instructions up from the load, to see if it is obviously live at
874 // the entry to its block.
875 BasicBlock::iterator BBIt = LI;
877 if (Value *AvailableVal =
878 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
879 // If the value if the load is locally available within the block, just use
880 // it. This frequently occurs for reg2mem'd allocas.
881 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
883 // If the returned value is the load itself, replace with an undef. This can
884 // only happen in dead loops.
885 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
886 if (AvailableVal->getType() != LI->getType())
888 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
889 LI->replaceAllUsesWith(AvailableVal);
890 LI->eraseFromParent();
894 // Otherwise, if we scanned the whole block and got to the top of the block,
895 // we know the block is locally transparent to the load. If not, something
896 // might clobber its value.
897 if (BBIt != LoadBB->begin())
900 // If all of the loads and stores that feed the value have the same AA tags,
901 // then we can propagate them onto any newly inserted loads.
903 LI->getAAMetadata(AATags);
905 SmallPtrSet<BasicBlock*, 8> PredsScanned;
906 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
907 AvailablePredsTy AvailablePreds;
908 BasicBlock *OneUnavailablePred = nullptr;
910 // If we got here, the loaded value is transparent through to the start of the
911 // block. Check to see if it is available in any of the predecessor blocks.
912 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
914 BasicBlock *PredBB = *PI;
916 // If we already scanned this predecessor, skip it.
917 if (!PredsScanned.insert(PredBB).second)
920 // Scan the predecessor to see if the value is available in the pred.
921 BBIt = PredBB->end();
922 AAMDNodes ThisAATags;
923 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
924 nullptr, &ThisAATags);
925 if (!PredAvailable) {
926 OneUnavailablePred = PredBB;
930 // If AA tags disagree or are not present, forget about them.
931 if (AATags != ThisAATags) AATags = AAMDNodes();
933 // If so, this load is partially redundant. Remember this info so that we
934 // can create a PHI node.
935 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
938 // If the loaded value isn't available in any predecessor, it isn't partially
940 if (AvailablePreds.empty()) return false;
942 // Okay, the loaded value is available in at least one (and maybe all!)
943 // predecessors. If the value is unavailable in more than one unique
944 // predecessor, we want to insert a merge block for those common predecessors.
945 // This ensures that we only have to insert one reload, thus not increasing
947 BasicBlock *UnavailablePred = nullptr;
949 // If there is exactly one predecessor where the value is unavailable, the
950 // already computed 'OneUnavailablePred' block is it. If it ends in an
951 // unconditional branch, we know that it isn't a critical edge.
952 if (PredsScanned.size() == AvailablePreds.size()+1 &&
953 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
954 UnavailablePred = OneUnavailablePred;
955 } else if (PredsScanned.size() != AvailablePreds.size()) {
956 // Otherwise, we had multiple unavailable predecessors or we had a critical
957 // edge from the one.
958 SmallVector<BasicBlock*, 8> PredsToSplit;
959 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
961 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
962 AvailablePredSet.insert(AvailablePreds[i].first);
964 // Add all the unavailable predecessors to the PredsToSplit list.
965 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
968 // If the predecessor is an indirect goto, we can't split the edge.
969 if (isa<IndirectBrInst>(P->getTerminator()))
972 if (!AvailablePredSet.count(P))
973 PredsToSplit.push_back(P);
976 // Split them out to their own block.
978 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split");
981 // If the value isn't available in all predecessors, then there will be
982 // exactly one where it isn't available. Insert a load on that edge and add
983 // it to the AvailablePreds list.
984 if (UnavailablePred) {
985 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
986 "Can't handle critical edge here!");
987 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
989 UnavailablePred->getTerminator());
990 NewVal->setDebugLoc(LI->getDebugLoc());
992 NewVal->setAAMetadata(AATags);
994 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
997 // Now we know that each predecessor of this block has a value in
998 // AvailablePreds, sort them for efficient access as we're walking the preds.
999 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1001 // Create a PHI node at the start of the block for the PRE'd load value.
1002 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1003 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1006 PN->setDebugLoc(LI->getDebugLoc());
1008 // Insert new entries into the PHI for each predecessor. A single block may
1009 // have multiple entries here.
1010 for (pred_iterator PI = PB; PI != PE; ++PI) {
1011 BasicBlock *P = *PI;
1012 AvailablePredsTy::iterator I =
1013 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1014 std::make_pair(P, (Value*)nullptr));
1016 assert(I != AvailablePreds.end() && I->first == P &&
1017 "Didn't find entry for predecessor!");
1019 // If we have an available predecessor but it requires casting, insert the
1020 // cast in the predecessor and use the cast. Note that we have to update the
1021 // AvailablePreds vector as we go so that all of the PHI entries for this
1022 // predecessor use the same bitcast.
1023 Value *&PredV = I->second;
1024 if (PredV->getType() != LI->getType())
1025 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1026 P->getTerminator());
1028 PN->addIncoming(PredV, I->first);
1031 //cerr << "PRE: " << *LI << *PN << "\n";
1033 LI->replaceAllUsesWith(PN);
1034 LI->eraseFromParent();
1039 /// FindMostPopularDest - The specified list contains multiple possible
1040 /// threadable destinations. Pick the one that occurs the most frequently in
1043 FindMostPopularDest(BasicBlock *BB,
1044 const SmallVectorImpl<std::pair<BasicBlock*,
1045 BasicBlock*> > &PredToDestList) {
1046 assert(!PredToDestList.empty());
1048 // Determine popularity. If there are multiple possible destinations, we
1049 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1050 // blocks with known and real destinations to threading undef. We'll handle
1051 // them later if interesting.
1052 DenseMap<BasicBlock*, unsigned> DestPopularity;
1053 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1054 if (PredToDestList[i].second)
1055 DestPopularity[PredToDestList[i].second]++;
1057 // Find the most popular dest.
1058 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1059 BasicBlock *MostPopularDest = DPI->first;
1060 unsigned Popularity = DPI->second;
1061 SmallVector<BasicBlock*, 4> SamePopularity;
1063 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1064 // If the popularity of this entry isn't higher than the popularity we've
1065 // seen so far, ignore it.
1066 if (DPI->second < Popularity)
1068 else if (DPI->second == Popularity) {
1069 // If it is the same as what we've seen so far, keep track of it.
1070 SamePopularity.push_back(DPI->first);
1072 // If it is more popular, remember it.
1073 SamePopularity.clear();
1074 MostPopularDest = DPI->first;
1075 Popularity = DPI->second;
1079 // Okay, now we know the most popular destination. If there is more than one
1080 // destination, we need to determine one. This is arbitrary, but we need
1081 // to make a deterministic decision. Pick the first one that appears in the
1083 if (!SamePopularity.empty()) {
1084 SamePopularity.push_back(MostPopularDest);
1085 TerminatorInst *TI = BB->getTerminator();
1086 for (unsigned i = 0; ; ++i) {
1087 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1089 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1090 TI->getSuccessor(i)) == SamePopularity.end())
1093 MostPopularDest = TI->getSuccessor(i);
1098 // Okay, we have finally picked the most popular destination.
1099 return MostPopularDest;
1102 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1103 ConstantPreference Preference,
1104 Instruction *CxtI) {
1105 // If threading this would thread across a loop header, don't even try to
1107 if (LoopHeaders.count(BB))
1110 PredValueInfoTy PredValues;
1111 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1114 assert(!PredValues.empty() &&
1115 "ComputeValueKnownInPredecessors returned true with no values");
1117 DEBUG(dbgs() << "IN BB: " << *BB;
1118 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1119 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1120 << *PredValues[i].first
1121 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1124 // Decide what we want to thread through. Convert our list of known values to
1125 // a list of known destinations for each pred. This also discards duplicate
1126 // predecessors and keeps track of the undefined inputs (which are represented
1127 // as a null dest in the PredToDestList).
1128 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1129 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1131 BasicBlock *OnlyDest = nullptr;
1132 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1134 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1135 BasicBlock *Pred = PredValues[i].second;
1136 if (!SeenPreds.insert(Pred).second)
1137 continue; // Duplicate predecessor entry.
1139 // If the predecessor ends with an indirect goto, we can't change its
1141 if (isa<IndirectBrInst>(Pred->getTerminator()))
1144 Constant *Val = PredValues[i].first;
1147 if (isa<UndefValue>(Val))
1149 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1150 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1151 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1152 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1154 assert(isa<IndirectBrInst>(BB->getTerminator())
1155 && "Unexpected terminator");
1156 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1159 // If we have exactly one destination, remember it for efficiency below.
1160 if (PredToDestList.empty())
1162 else if (OnlyDest != DestBB)
1163 OnlyDest = MultipleDestSentinel;
1165 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1168 // If all edges were unthreadable, we fail.
1169 if (PredToDestList.empty())
1172 // Determine which is the most common successor. If we have many inputs and
1173 // this block is a switch, we want to start by threading the batch that goes
1174 // to the most popular destination first. If we only know about one
1175 // threadable destination (the common case) we can avoid this.
1176 BasicBlock *MostPopularDest = OnlyDest;
1178 if (MostPopularDest == MultipleDestSentinel)
1179 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1181 // Now that we know what the most popular destination is, factor all
1182 // predecessors that will jump to it into a single predecessor.
1183 SmallVector<BasicBlock*, 16> PredsToFactor;
1184 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1185 if (PredToDestList[i].second == MostPopularDest) {
1186 BasicBlock *Pred = PredToDestList[i].first;
1188 // This predecessor may be a switch or something else that has multiple
1189 // edges to the block. Factor each of these edges by listing them
1190 // according to # occurrences in PredsToFactor.
1191 TerminatorInst *PredTI = Pred->getTerminator();
1192 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1193 if (PredTI->getSuccessor(i) == BB)
1194 PredsToFactor.push_back(Pred);
1197 // If the threadable edges are branching on an undefined value, we get to pick
1198 // the destination that these predecessors should get to.
1199 if (!MostPopularDest)
1200 MostPopularDest = BB->getTerminator()->
1201 getSuccessor(GetBestDestForJumpOnUndef(BB));
1203 // Ok, try to thread it!
1204 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1207 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1208 /// a PHI node in the current block. See if there are any simplifications we
1209 /// can do based on inputs to the phi node.
1211 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1212 BasicBlock *BB = PN->getParent();
1214 // TODO: We could make use of this to do it once for blocks with common PHI
1216 SmallVector<BasicBlock*, 1> PredBBs;
1219 // If any of the predecessor blocks end in an unconditional branch, we can
1220 // *duplicate* the conditional branch into that block in order to further
1221 // encourage jump threading and to eliminate cases where we have branch on a
1222 // phi of an icmp (branch on icmp is much better).
1223 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1224 BasicBlock *PredBB = PN->getIncomingBlock(i);
1225 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1226 if (PredBr->isUnconditional()) {
1227 PredBBs[0] = PredBB;
1228 // Try to duplicate BB into PredBB.
1229 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1237 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1238 /// a xor instruction in the current block. See if there are any
1239 /// simplifications we can do based on inputs to the xor.
1241 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1242 BasicBlock *BB = BO->getParent();
1244 // If either the LHS or RHS of the xor is a constant, don't do this
1246 if (isa<ConstantInt>(BO->getOperand(0)) ||
1247 isa<ConstantInt>(BO->getOperand(1)))
1250 // If the first instruction in BB isn't a phi, we won't be able to infer
1251 // anything special about any particular predecessor.
1252 if (!isa<PHINode>(BB->front()))
1255 // If we have a xor as the branch input to this block, and we know that the
1256 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1257 // the condition into the predecessor and fix that value to true, saving some
1258 // logical ops on that path and encouraging other paths to simplify.
1260 // This copies something like this:
1263 // %X = phi i1 [1], [%X']
1264 // %Y = icmp eq i32 %A, %B
1265 // %Z = xor i1 %X, %Y
1270 // %Y = icmp ne i32 %A, %B
1273 PredValueInfoTy XorOpValues;
1275 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1277 assert(XorOpValues.empty());
1278 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1284 assert(!XorOpValues.empty() &&
1285 "ComputeValueKnownInPredecessors returned true with no values");
1287 // Scan the information to see which is most popular: true or false. The
1288 // predecessors can be of the set true, false, or undef.
1289 unsigned NumTrue = 0, NumFalse = 0;
1290 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1291 if (isa<UndefValue>(XorOpValues[i].first))
1292 // Ignore undefs for the count.
1294 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1300 // Determine which value to split on, true, false, or undef if neither.
1301 ConstantInt *SplitVal = nullptr;
1302 if (NumTrue > NumFalse)
1303 SplitVal = ConstantInt::getTrue(BB->getContext());
1304 else if (NumTrue != 0 || NumFalse != 0)
1305 SplitVal = ConstantInt::getFalse(BB->getContext());
1307 // Collect all of the blocks that this can be folded into so that we can
1308 // factor this once and clone it once.
1309 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1310 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1311 if (XorOpValues[i].first != SplitVal &&
1312 !isa<UndefValue>(XorOpValues[i].first))
1315 BlocksToFoldInto.push_back(XorOpValues[i].second);
1318 // If we inferred a value for all of the predecessors, then duplication won't
1319 // help us. However, we can just replace the LHS or RHS with the constant.
1320 if (BlocksToFoldInto.size() ==
1321 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1323 // If all preds provide undef, just nuke the xor, because it is undef too.
1324 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1325 BO->eraseFromParent();
1326 } else if (SplitVal->isZero()) {
1327 // If all preds provide 0, replace the xor with the other input.
1328 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1329 BO->eraseFromParent();
1331 // If all preds provide 1, set the computed value to 1.
1332 BO->setOperand(!isLHS, SplitVal);
1338 // Try to duplicate BB into PredBB.
1339 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1343 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1344 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1345 /// NewPred using the entries from OldPred (suitably mapped).
1346 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1347 BasicBlock *OldPred,
1348 BasicBlock *NewPred,
1349 DenseMap<Instruction*, Value*> &ValueMap) {
1350 for (BasicBlock::iterator PNI = PHIBB->begin();
1351 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1352 // Ok, we have a PHI node. Figure out what the incoming value was for the
1354 Value *IV = PN->getIncomingValueForBlock(OldPred);
1356 // Remap the value if necessary.
1357 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1358 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1359 if (I != ValueMap.end())
1363 PN->addIncoming(IV, NewPred);
1367 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1368 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1369 /// across BB. Transform the IR to reflect this change.
1370 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1371 const SmallVectorImpl<BasicBlock*> &PredBBs,
1372 BasicBlock *SuccBB) {
1373 // If threading to the same block as we come from, we would infinite loop.
1375 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1376 << "' - would thread to self!\n");
1380 // If threading this would thread across a loop header, don't thread the edge.
1381 // See the comments above FindLoopHeaders for justifications and caveats.
1382 if (LoopHeaders.count(BB)) {
1383 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1384 << "' to dest BB '" << SuccBB->getName()
1385 << "' - it might create an irreducible loop!\n");
1389 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1390 if (JumpThreadCost > BBDupThreshold) {
1391 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1392 << "' - Cost is too high: " << JumpThreadCost << "\n");
1396 // And finally, do it! Start by factoring the predecessors is needed.
1398 if (PredBBs.size() == 1)
1399 PredBB = PredBBs[0];
1401 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1402 << " common predecessors.\n");
1403 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm");
1406 // And finally, do it!
1407 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1408 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1409 << ", across block:\n "
1412 LVI->threadEdge(PredBB, BB, SuccBB);
1414 // We are going to have to map operands from the original BB block to the new
1415 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1416 // account for entry from PredBB.
1417 DenseMap<Instruction*, Value*> ValueMapping;
1419 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1420 BB->getName()+".thread",
1421 BB->getParent(), BB);
1422 NewBB->moveAfter(PredBB);
1424 BasicBlock::iterator BI = BB->begin();
1425 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1426 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1428 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1429 // mapping and using it to remap operands in the cloned instructions.
1430 for (; !isa<TerminatorInst>(BI); ++BI) {
1431 Instruction *New = BI->clone();
1432 New->setName(BI->getName());
1433 NewBB->getInstList().push_back(New);
1434 ValueMapping[BI] = New;
1436 // Remap operands to patch up intra-block references.
1437 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1438 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1439 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1440 if (I != ValueMapping.end())
1441 New->setOperand(i, I->second);
1445 // We didn't copy the terminator from BB over to NewBB, because there is now
1446 // an unconditional jump to SuccBB. Insert the unconditional jump.
1447 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1448 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1450 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1451 // PHI nodes for NewBB now.
1452 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1454 // If there were values defined in BB that are used outside the block, then we
1455 // now have to update all uses of the value to use either the original value,
1456 // the cloned value, or some PHI derived value. This can require arbitrary
1457 // PHI insertion, of which we are prepared to do, clean these up now.
1458 SSAUpdater SSAUpdate;
1459 SmallVector<Use*, 16> UsesToRename;
1460 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1461 // Scan all uses of this instruction to see if it is used outside of its
1462 // block, and if so, record them in UsesToRename.
1463 for (Use &U : I->uses()) {
1464 Instruction *User = cast<Instruction>(U.getUser());
1465 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1466 if (UserPN->getIncomingBlock(U) == BB)
1468 } else if (User->getParent() == BB)
1471 UsesToRename.push_back(&U);
1474 // If there are no uses outside the block, we're done with this instruction.
1475 if (UsesToRename.empty())
1478 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1480 // We found a use of I outside of BB. Rename all uses of I that are outside
1481 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1482 // with the two values we know.
1483 SSAUpdate.Initialize(I->getType(), I->getName());
1484 SSAUpdate.AddAvailableValue(BB, I);
1485 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1487 while (!UsesToRename.empty())
1488 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1489 DEBUG(dbgs() << "\n");
1493 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1494 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1495 // us to simplify any PHI nodes in BB.
1496 TerminatorInst *PredTerm = PredBB->getTerminator();
1497 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1498 if (PredTerm->getSuccessor(i) == BB) {
1499 BB->removePredecessor(PredBB, true);
1500 PredTerm->setSuccessor(i, NewBB);
1503 // At this point, the IR is fully up to date and consistent. Do a quick scan
1504 // over the new instructions and zap any that are constants or dead. This
1505 // frequently happens because of phi translation.
1506 SimplifyInstructionsInBlock(NewBB, TLI);
1508 // Threaded an edge!
1513 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1514 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1515 /// If we can duplicate the contents of BB up into PredBB do so now, this
1516 /// improves the odds that the branch will be on an analyzable instruction like
1518 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1519 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1520 assert(!PredBBs.empty() && "Can't handle an empty set");
1522 // If BB is a loop header, then duplicating this block outside the loop would
1523 // cause us to transform this into an irreducible loop, don't do this.
1524 // See the comments above FindLoopHeaders for justifications and caveats.
1525 if (LoopHeaders.count(BB)) {
1526 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1527 << "' into predecessor block '" << PredBBs[0]->getName()
1528 << "' - it might create an irreducible loop!\n");
1532 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1533 if (DuplicationCost > BBDupThreshold) {
1534 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1535 << "' - Cost is too high: " << DuplicationCost << "\n");
1539 // And finally, do it! Start by factoring the predecessors is needed.
1541 if (PredBBs.size() == 1)
1542 PredBB = PredBBs[0];
1544 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1545 << " common predecessors.\n");
1546 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm");
1549 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1551 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1552 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1553 << DuplicationCost << " block is:" << *BB << "\n");
1555 // Unless PredBB ends with an unconditional branch, split the edge so that we
1556 // can just clone the bits from BB into the end of the new PredBB.
1557 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1559 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1560 PredBB = SplitEdge(PredBB, BB);
1561 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1564 // We are going to have to map operands from the original BB block into the
1565 // PredBB block. Evaluate PHI nodes in BB.
1566 DenseMap<Instruction*, Value*> ValueMapping;
1568 BasicBlock::iterator BI = BB->begin();
1569 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1570 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1571 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1572 // mapping and using it to remap operands in the cloned instructions.
1573 for (; BI != BB->end(); ++BI) {
1574 Instruction *New = BI->clone();
1576 // Remap operands to patch up intra-block references.
1577 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1578 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1579 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1580 if (I != ValueMapping.end())
1581 New->setOperand(i, I->second);
1584 // If this instruction can be simplified after the operands are updated,
1585 // just use the simplified value instead. This frequently happens due to
1588 SimplifyInstruction(New, BB->getModule()->getDataLayout())) {
1590 ValueMapping[BI] = IV;
1592 // Otherwise, insert the new instruction into the block.
1593 New->setName(BI->getName());
1594 PredBB->getInstList().insert(OldPredBranch, New);
1595 ValueMapping[BI] = New;
1599 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1600 // add entries to the PHI nodes for branch from PredBB now.
1601 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1602 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1604 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1607 // If there were values defined in BB that are used outside the block, then we
1608 // now have to update all uses of the value to use either the original value,
1609 // the cloned value, or some PHI derived value. This can require arbitrary
1610 // PHI insertion, of which we are prepared to do, clean these up now.
1611 SSAUpdater SSAUpdate;
1612 SmallVector<Use*, 16> UsesToRename;
1613 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1614 // Scan all uses of this instruction to see if it is used outside of its
1615 // block, and if so, record them in UsesToRename.
1616 for (Use &U : I->uses()) {
1617 Instruction *User = cast<Instruction>(U.getUser());
1618 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1619 if (UserPN->getIncomingBlock(U) == BB)
1621 } else if (User->getParent() == BB)
1624 UsesToRename.push_back(&U);
1627 // If there are no uses outside the block, we're done with this instruction.
1628 if (UsesToRename.empty())
1631 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1633 // We found a use of I outside of BB. Rename all uses of I that are outside
1634 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1635 // with the two values we know.
1636 SSAUpdate.Initialize(I->getType(), I->getName());
1637 SSAUpdate.AddAvailableValue(BB, I);
1638 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1640 while (!UsesToRename.empty())
1641 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1642 DEBUG(dbgs() << "\n");
1645 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1647 BB->removePredecessor(PredBB, true);
1649 // Remove the unconditional branch at the end of the PredBB block.
1650 OldPredBranch->eraseFromParent();
1656 /// TryToUnfoldSelect - Look for blocks of the form
1662 /// %p = phi [%a, %bb] ...
1666 /// And expand the select into a branch structure if one of its arms allows %c
1667 /// to be folded. This later enables threading from bb1 over bb2.
1668 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1669 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1670 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1671 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1673 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1674 CondLHS->getParent() != BB)
1677 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1678 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1679 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1681 // Look if one of the incoming values is a select in the corresponding
1683 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1686 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1687 if (!PredTerm || !PredTerm->isUnconditional())
1690 // Now check if one of the select values would allow us to constant fold the
1691 // terminator in BB. We don't do the transform if both sides fold, those
1692 // cases will be threaded in any case.
1693 LazyValueInfo::Tristate LHSFolds =
1694 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1695 CondRHS, Pred, BB, CondCmp);
1696 LazyValueInfo::Tristate RHSFolds =
1697 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1698 CondRHS, Pred, BB, CondCmp);
1699 if ((LHSFolds != LazyValueInfo::Unknown ||
1700 RHSFolds != LazyValueInfo::Unknown) &&
1701 LHSFolds != RHSFolds) {
1702 // Expand the select.
1711 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1712 BB->getParent(), BB);
1713 // Move the unconditional branch to NewBB.
1714 PredTerm->removeFromParent();
1715 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1716 // Create a conditional branch and update PHI nodes.
1717 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1718 CondLHS->setIncomingValue(I, SI->getFalseValue());
1719 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1720 // The select is now dead.
1721 SI->eraseFromParent();
1723 // Update any other PHI nodes in BB.
1724 for (BasicBlock::iterator BI = BB->begin();
1725 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1727 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);